Machines of Love & Grace – Building Truly Inclusive AI
summary
Argues that bias lives in the training data and not just the model, citing 1.6 million Enron emails used for NLP, Princeton word-embedding findings that equate programmer with man and homemaker with woman, and the fact that only 13% of Wikipedia editors identify as women; describes tools and fixes: a Crescendo Slack bias detector that senses unconscious bias and delivers personalized inclusion training, Ethics Net curating pro-social datasets, copyright law as a lever to rebalance training corpora, and the Correlates of War dataset’s undergraduate coding errors as a caution.
AI-written summary · about
chapters (5)
- 0:00Welcome
- 3:40AI inclusivity challenges
- 6:33Bias in open-source datasets
- 8:59Human-centered data coding
- 10:47Q&A
related talks
34m(Un)Ethical Machines? Why AI Bias Is a Problem and What to Do About It
38mSocial Implications of Bias in Machine Learning
30mInside AI: What Every Professional Needs to Know to Stay Relevant, Protected, and Ahead
56mAn AI with an Agenda: How Our Cognitive Biases Leak Into Machine Learning
28mBreaking the Chains of Oppressive Software
30mFighting Prejudice in Artificial Intelligence
42mTowards Inclusive AI: Understanding bias in ML models
05mToward Ethical AI